Decentralisation, clientelism and social protection programmes: a study of India’s MGNREGA

Abstract Does decentralisation promote clientelism? If yes, through which mechanisms? We answer these questions through an analysis of India’s (and the world’s) largest workfare programme, the Mahatma Gandhi National Rural Employment Guarantee Act (MGNREGA), in two Indian states: Rajasthan and Andhra Pradesh (AP). The two states adopted radically different implementation models: Rajasthan’s decentralised one stands in contrast with Andhra Pradesh’s centralised and bureaucracy-led model. Using a mixed method approach, we find that in both states local implementers have incentives to distribute MGNREGA work in a clientelistic fashion. However, in Rajasthan, these incentives are stronger, because of the decentralised implementation model. Accordingly, our quantitative evidence shows that clientelism is more serious a problem in Rajasthan than in AP.


Introduction
The world's largest workfare programme, India's Mahatma Gandhi National Rural Employment Guarantee Act (MGNREGA), has been adopted via two radically different implementation models in two Indian states, Andhra Pradesh (AP) 1 and Rajasthan. AP implemented the programme through the state bureaucracy, while Rajasthan devolved implementation to elected village councils (the gram panchayats, or GPs). A qualitative and quantitative analysis of data from each state allows us to ask a number of questions. Does democratic decentralisation promote clientelism? If so, which model is more effective at reducing clientelism? Furthermore, is AP's 'depoliticised' implementation less prone to capturing by local elites?
We find that some degree of clientelism in the allocation of MGNREGA work is present in both states. However, clientelism is significantly less pronounced in AP's centralised setting. This is explained by the different incentives generated by the different institutional models: in AP, local implementers have stronger incentives to maximise the generation of MGNREGA employment, which in turn incentivises them to distribute work more broadly, going beyond their circle of supporters, relatives and friends. In Rajasthan's decentralised model, these incentives are missing and local implementers distribute work 'politically' in order to build support for themselves or their parties (see Section 2 below). This paper makes three main contributions. First, we add to the literature on decentralisation and governance, which has 'largely ignored' the effects of the former on the latter (Faguet, 2014, p. 10). This is rather ironic, considering that one of the main objectives of the decentralisation wave of the 1980s and 1990s was precisely to promote good governance (Manor, 1999;WDR, 2004;Widmalm, 2008). As clientelism has several negative consequences on governance, 2 we contribute to this literature not only by establishing a correlation between decentralised implementation and clientelism, but also by looking inside the 'black box' (Hedström & Ylikoski, 2010) and explaining the mechanisms through which clientelism finds a more or less prominent role in the implementation of the MGNREGA.
Secondly, we provide an answer to a seriously under-researched question, namely whether decentralisation promotes clientelism (Bardhan & Mookherjee, 2016). Existing studies reach mixed conclusions: some authors conclude that decentralisation does reduce the scope for clientelism and capture, 3 at least under certain conditions (Bardhan, Mitra, Mookherjee, & Sarkar, 2015;Faguet & Pöschl, 2015), while others find that decentralisation actually promotes it (Crook, 2003;Manor, 1999;Sadanandan, 2012). We look at the relationship between decentralisation and clientelism from a particularly advantageous point of view: by comparing two states that display similar factors associated with clientelism in the literature, we can focus on the hypothesised institutional driver of clientelism, namely democratic decentralisation. Our paper also adds to the literature on the variance of clientelistic practices at the subnational level. 4 Third, building on what Fox (1994) refers to as 'semiclientelism' and on recent scholarship (Hicken, 2011), we argue that clientelism should not be seen as a dichotomous category (the absence or presence of it), but rather as a continuous variable (from a purely programmatic to a purely clientelistic distribution). This is because, as detailed below, clientelism is the result of a complex set of incentives, depending in turn on the institutional set up in which implementation takes place, as well as on the broader political and social context. In other words, clientelism is not an exclusive relationship between a patron and a client that interacts in a vacuum, but it depends also on other actors' interests, power and agency.
The paper is structured as follows. The next section details our research design. Section 3 describes the functioning of the MGNREGA in our two fieldwork states and presents the basic premises upon which the paper relies. Section 4 unpacks the different sets of incentives to which MGNREGA local implementers are subjected in the two states. Qualitative evidence is corroborated, in Section 5, by a quantitative analysis of the role of GP-level implementers in the allocation of MGNREGA work. Section 6 concludes.

Research design and methodology
The MGNREGA guarantees 100 days of employment per year in public works to every rural household. Rajasthan and AP display similar factors usually associated with clientelism, such as: the electoral system (McGillivray, 2004); levels of poverty and of party competition (Brusco, Nazareno, & Stokes, 2004;Kitschelt & Wilkinson, 2007;Sadanandan, 2012;Weitz-Shapiro, 2014); and the role of the state in the economy and the size of the informal sector (Bardhan & Mookherjee, 2016;Chandra, 2004). Furthermore, the design and quality of institutions is similar (Golden, 2003;Heath & Tillin, 2018), as indicated by the fact that both states were quite successful at implementing the MGNREGA policy ( Table 1) 5 The two states, however, adopted radically different models of implementation. In Rajasthan, the key implementer at the local level is the sarpanch, who is the head of the elected council (GP) and is therefore accountable to the beneficiaries who form a sizeable part of the electorate. In AP, the GPs are excluded from playing any meaningful role in the MGNREGA and the key decision-maker is an employee of the state government, the Field Assistant (FA), who is accountable to the higher echelons of the state administration (Jenkins & Manor, 2017, p. 255;Maiorano, 2014;Veeraraghavan, 2015, p. 25 and p. 34). 6 This was a conscious choice of the state administration. The architects of the MGNREGA in AP shared an 'ambedkarite' view of the GPs. 7 One of the top officials in the Rural Development Department of AP told us: We decided not to involve the GPs because 95 per cent of them are dominated by the upper castes and they rule according to feudal rules. We thought that nothing good for the poor could come out of them. (interview, Hyderabad, 20/12/2012) Other officials expressed similar feelings of mistrust: one of them told us that they feared that 'once you open the gate [of decentralised implementation] you are not able to control it' (interview, Hyderabad, 17 December 2012); another very senior official stated that the MGNREGA could just 'not work' if implemented through the GPs: 'they don't even have computers!' (interview, Hyderabad, 15 December 2012).
In other words, the two states exemplify a major debate in development research: Rajasthan's decentralised and GP-led model stands in contrast with AP's centralised, technocratic and administration-led model. This provides an empirical basis to ask, which model is less prone to clientelism and why?
We used a broad definition of clientelism as an exchange system where state officials (elected or appointed) discretionally distribute state resources to citizens in exchange for political support. By 'political support' we do not mean voting alone, but also other forms of support that increase the power and status of the official, like favours, services, and respect, all very valuable resources in an Indian village.
We adopt a mixed method approach combining 150 semi-structured interviews with survey data collected between October 2013 and January 2014, in six districts (three in each state) chosen to maximise intra-state variety. 8 In each district, two GPs were randomly selected within the same sub-district 9 and respondents were randomly selected among the GP's MGNREGA beneficiaries. 10 The next section spells out three important premises on which the rest of the paper is based.

MGNREGA: a post-clientelistic policy?
By conferring the right to obtain employment on demand, the MGNREGA was specifically designed as a 'post-clientelistic' policy (Elliott, 2011;Manor, 2013) and some studies do find very limited distributive distortions (Johnson, 2009;Sheahan, Liu, Barrett, & Narayanan, 2014). However, numerous studies show that supply-side issues -like administrative incapacity and lack of political commitment -result in work being 'rationed' (Chopra, 2014(Chopra, , 2015Dutta, Murgai, Ravallion, & van de Walle, 2014;Jenkins & Manor, 2017;Mukherji & Jha, 2014;Ravi & Engler, 2009) and that clientelism affects implementation (Das, 2015;Khosla, 2011;Mukhopadhyay, Himanshu, & Sharan, 2015). What these studies do not do, however, is explain why clientelism is present or absent from the implementation of the programme. Our paper provides such an explanation through an analysis of the mechanisms that shape the distribution of MGNREGA work. This paper is based on three premises. First, the more MGNREGA work is rationed, the more local implementers will have the incentive to distribute it in a clientelistic fashion, as scarcity is a key determinant of a distorted distribution (Chubb, 1982;Weitz-Shapiro, 2014). Das and Maiorano (2015) find this to be true in the context of MGNREGA. From the perspective of a local implementer, a greater availability of MGNREGA work makes it less costly to distribute to someone who will not provide (or is less likely to provide) political support in return. Secondly, local implementers have the power to determine how scarce MGNREGA work is in their GP. It should be noted that there are a number of supply-side factors that determine the scarcity of MGNREGA work that are beyond the control of local implementers, the most important of which is funding from the central (federal) government, which provides 90% of the total funds. Figure 1 shows that the central government has significantly reduced funding for the programme over the last few years.
Limited funding is a significant constraint to the generation of MGNREGA employment, as it limits the number of worksites that can be opened in any given GP, which in turn limits the number of people that can be employed. In principle, the central government should provide enough financial resources to cover the theoretical possibility that all rural households demand 100 days of employment in any given year. However, the central government simply violates this provision of the Act and instead caps budgetary allocations. In other words, if the central government does not allocate enough resources, local implementers will not be able to meet the demand for MGNREGA work.
Given the constraints on employment generation determined by central funding and other supply-side issues, local implementers can nevertheless determine how scarce (or abundant) MGNREGA work actually is in their GP. This is so because they have the monopoly of information on all aspects of the programme, including how many and which projects are approved by sub-district level officials. Therefore, the sarpanch in Rajasthan and the FA in AP can determine levels of scarcity by controlling the number and the type (more or less labour intensive) of worksites that are started in their GP as no worksite can start and no one can be employed without their sanction. In the next section we analyse the incentives that local implementers have to generate more (or less) employment in their GPs.
The third premise is that local implementers have the power to distribute MGNREGA work at their discretion, which is a hallmark of clientelism. Our qualitative evidence fully supports this, and so does the literature (Dunning & Nilekani, 2013;. In Rajasthan, the sarpanches have a crucial influence on the selection of beneficiaries . Our survey shows that 41% of the respondents approach the sarpanch to demand work and 66% of them think that he/ she is the most-important decision-maker within the programme.
A sarpanch reported: When a worksite is approved, I approach the people living nearby and I ask if they want to work. I hired a person (that I pay from my own pocket) to help them fill the 'Form 6' . 11 (Interview, Kaurali disitrct, February 7, 2014) A rozgar sevak 12 described a similar procedure: When the sarpanch tells me that a worksite is ready to be opened, I approach the workers that live near the worksite location and ask them if they want to work. If they do, I fill in the 'Form 6s' and submit them to the block office. (Interview, Sirohi, district February 10, 2014) In short, work is available only if and when the sarpanch offers it, as they have full authority over the list and location of works to be taken up. 13 In AP the institutional set up is very different. The key implementer at the GP level is the FA, a contractual employee of the state government, appointed at the GP level. However, they are not accountable to the GP, but to the mandal (sub-district) level officials. Moreover, as we shall see below, FAs are also closely monitored by Members of the Legislative Assembly (MLAs). Like the sarpanches in Rajasthan, the FA controls every aspect of the programme including the selection of the beneficiaries and, to a lesser extent, the worksite location (interview, state official, Hyderabad, December 6, 2012).
One 'mate' told us: 14 When some works are ready to start, the FA calls me and asks if we are interested in that particular type of work. I then ask to my group's members and report to the FA. We are lucky because I am in a good relationship [sic] with the FA, so he calls me often. (interview, Chittoor district, October 4, 2013) In other cases, the FA decided to allocate work on a rotation basis to avoid discrimination and protest from the villagers (interview with two FAs, Chittoor and Karimnagar districts, October 9 and 31, 2013). In other cases, the opening of a worksite is announced with drums, so that people interested in working come to the GP office and let the FA know (direct observation, Karimnagar district, November 2, 2013).
In short, despite the difference in institutional settings, the situation is strongly reminiscent of Rajasthan: work is provided only if and when the FA decides so. 15 In the words of a state official, 'if the FA doesn't want to provide jobs, he doesn't' (interview, Hyderabad, December 19, 2012). Our survey data show that 69.2% of the respondents are 'offered' work (rather than 'demanded') and 69% indicate that the FA is the most important decision-maker in MGNREGA.
To sum up, the MGNREGA is implemented through quite different institutional set ups in the two states. In Rajasthan, the key implementer (the sarpanch) is directly accountable to the beneficiaries who form a sizeable part of the electorate, while in AP the process is completely in the hands of the state's administration, which is hardly accountable to the MGNREGA beneficiaries. Despite the different institutional set ups, however, the process of allocation of MGNREGA jobs is similar in the two states. In both cases, it depends on the discretion of one key actor at the GP level. We now look at the incentives that local implementers have to generate as much (or as little) MGNREGA employment as possible.

Political incentives
Local implementers' political incentives for the maximisation of MGNREGA work are different in the two states, mainly because of the different institutional set up. The most obvious political incentive for providing MGNREGA jobs in (decentralised) Rajasthan is the fact that the sarpanch is able to distribute tangible benefits -employment -to the poor who, in many cases, form a sizable part of the electorate. 16 Assuming that the sarpanch's primary objective is to be re-elected, he/she will have the incentive to use the MGNREGA to keep their support base intact and/or to enlarge it. This should incentivise them to provide as much work as possible.
However, budgetary constraints and low administrative capacity seriously limit the availability of MGNREGA work, as illustrated above. Furthermore, sarpanches know that their prospects of being re-elected are very low, especially if their post is reserved for women or other disadvantaged communities 17 (Chattopadhyay & Duflo, 2004). Hence, the incentive to maximise the generation of MGNREGA work and enlarge the distribution of employment is limited. In this situation, we can expect sarpanches to prioritise their supporters, families and relatives, when allocating the available work. This is consistent with our survey data (see below), with our interviews with the sarpanches of the surveyed GPs, and with the theory on clientelism. According to the latter, politicians tend to target the core of their support base when allocating scarce goods and services, especially when they have direct relations with their clients (like the sarpanches have) (Cox, 2010). This meets the theoretical expectations of Stokes et al. 's (2013) 'broker-mediated theory' , according to which the lower one descends in the party hierarchy, the higher the incentive to target core (rather than swing) voters. This expectation is also backed by studies of the MGNREGA in other states (Das, 2015;Dey & Sen, 2016). 18 In AP, FAs have political incentives too. The FA posts are frequently found to be political posts controlled by the local MLA, who are able to exert significant control over the bureaucracy (Iyer & Mani, 2012). MLAs are usually not interested in who gets MGNREGA work, but they are interested in making sure that work is available in their constituencies. 19 It is not uncommon, for example, for them to exert pressure on the state administration (in particular at the district and mandal level) and on the FAs to generate enough employment to satisfy people's demand, but at the same time protect the interests of the farming community are maintained (see below). Moreover, FAs are fully embedded into the GP's political economy and thus could be taking the sarpanch's (or other powerful actor's) political needs into consideration when allocating MGNREGA work. Additionally, FAs have the incentive to be as generous as possible when allocating work in order to build up their reputation and increase their status and power.
In both states, local implementers have a strong counter-incentive to generate employment: big farmers' opposition to the MGNREGA (Jakimow, 2014;Veeraraghavan, 2015). According to a senior official of the Rural Development Department of AP, pressure from the farmers' lobby to limit the availability of MGNREGA work is the 'single most important reason why we fail to provide 100 days of work to whoever demands it' (interview, Hyderabad, 6 August 2013). Landowners claim that the scheme has caused an increase in wages (thus making farming unprofitable) 20 and that it has become difficult to find labourers during the agricultural peak season. 21 Moreover, dominant caste landlords resent the fact that the MGNREGA has contributed to the alteration of power relationships at the village level (Jakimow, 2014;Maiorano, Thapar-Björkert, & Blomkvist, 2016;Roy, 2014Roy, , 2015. As one farmer we talked to put it: the MGNREGA 'is part of the system of injustice that the government created against the farmers. It is perpetuating the disrespect that we are experiencing from the lower castes' (interview, Anantpur district, January 2017).
To sum up, political incentives in both states push in two opposing directions as GP-level implementers must reconcile the interests of two opposing constituencies: usually powerless workers pushing for more MGNREGA work on the one hand, and usually very powerful farmers pushing for less availability of work on the other. This limits their incentive to maximize employment.

Economic incentives
A second type of incentive to provide MGNREGA work relates to the possibility by local level implementers to extract an illicit fee out of it. This is, of course, a type of incentive. However, in the case of Rajasthan (and of politically ambitious FAs in AP) it is also a way to fund their electoral campaigns. In fact, the increased importance of the office since the introduction of the MGNREGA has also brought about increased electoral competition and higher electoral expenses. 22 The main way through which local implementers in both states can pocket some money is through bribes on the purchase of the materials needed for the execution of works. This is in fact where the bulk of corruption is to be found (Afridi & Iversen, 2013). However, the expenditure on materials cannot exceed 40% of the total in any given GP. Thus, the more employment a local implementer provides, the higher would be the (absolute) expenditure on materials, from which they can extract a fee.
However, this incentive to maximise MGNREGA employment is limited by the strong transparency and accountability measures that the MGNREGA contains. Not only are all data related to programme implementation -including the flow of funds -available online, but the act prescribes that social audits are held regularly in every single GP.
AP is the only state that has institutionalised social audits and implements them regularly through an independent society (Aiyar, Kapoor Metha, & Samji, 2009;Akella & Kidambi, 2007). Rajasthan, on the other hand, has conducted a number of extremely successful social audits through civil society organisations such as the Mazdoor Kisan Shakti Sangatan (MKSS); however, in villages that lack a strong network of civil society organisations their efficacy is rather limited (Jenkins & Manor, 2017). In any case, as one sarpanch puts it, stealing from the MGNREGA 'is a lot of work for little reward; and with a high probability of being caught' (interview, Churu district, 1 February, 2014). Another sarpanch agrees: 'why should I try to steal from the MGNREGA when the probability of being caught is so high and the money must be shared with a lot of other people?' (Interview, Sirohi district 24 February, 2014). This of course does not mean that sarpanches renounce to steal from the programme, but it is clear that the incentive to maximise employment generation in order to extract an illicit fee is limited by the difficulties in doing so.
The situation in AP is even more difficult for those who want to steal from the programme, as social audits are regularly conducted in all GPs, and they constitute a deterrent for FAs and other officials who wish to supplement their salary. As Aiyar and Mehta (2015) found, in order to steal from the MGNREGA in AP it is now necessary to build large corruption networks which, by involving numerous actors, increase the possibility of being caught and diminish the amount of resources that can be pocketed.
Therefore, the transparency mechanisms in place limit the possibility of extracting an illicit fee out of it. The reduction in corruption witnessed in recent years is an indication of the efficacy of these measures (Drèze, 2014). This acts in opposition to the incentive to maximise MGNREGA employment described above. Furthermore, it reinforces the incentive to provide work according to political considerations, as these remain the most tangible benefits for local implementers.

Administrative incentives
The third type of incentive for the maximisation of MGNREGA work is administrative. In Rajasthan, these incentives are virtually absent, since the decentralised setting chosen for the implementation of the MGNREGA leaves the sarpanches free to set their own targets.
In AP, on the other hand, where the FA is accountable to the state government, the latter has the power to impose targets from above. In fact, FAs must generate at least 15,000 workdays per year. If they fail, their contract may be terminated, or their salary reduced. This constitutes an important incentive to maximise MGNREGA employment. An FA told us that at times she has to 'chase' villagers to work under MGNREGA in order to meet the target (interview, Karimnagar district, August 23, 2013). In fact, all the FAs we talked to were very concerned about meeting the target. In certain cases, this was completely unrealistic given the small size of the GP where the FAs worked. In all the GPs that we visited, the administrative incentive to provide a fixed number of persondays worked as a potent incentive to maximise MGNREGA work, which in turn pushed for the broadening of the pool of beneficiaries and hence a reduced scope for the allocation of MGNREGA work along strict clientelistic lines.

Quantitative analysis
As we have seen, different institutional settings generate different types of incentives in the two states; however, these have similar effects: political and economic incentives to maximise MGNREGA employment are off-set by powerful counter-incentives to keep MGNREGA employment to a minimum.
As Table 2 summarises, the incentives of GP-level implementers to maximise the allocation of MGNREGA employment (subject to scarcity determined by factors beyond their control) are stronger in centralised AP than in decentralised Rajasthan. This should make MGNREGA work less scarce in AP and therefore less costly to distribute widely and not along strict clientelistic lines. In this section, we adopt a regression analysis to test this hypothesis.
(a) Variables (i) Outcome Variables-getting work, number of days of work and earnings from MGNREGA We use three main dependent variables for the econometric analysis: whether the household received work or not; the number of work days; and the earnings from MGNREGA for the year 2013-2014. These data are taken from the official MGNREGA website, and associated with the job-card numbers recorded during the survey. We categorise getting work as '1' or '0' depending on whether the household received work or not, respectively. For the number of workdays and for earnings, we add 1 and take the logarithmic value of both these variables as the dependent variables to generate a value of zero for those who did not work.
(ii) Explanatory Variables Our primary variable of interest is the local implementers' political interests in the two states. The proxy used for political interests in Rajasthan is a dummy variable, which takes the value of '1' if the household resides in the same village (within the same GP) of the sarpanch, and the value of '0' otherwise. Many Indian GPs consist of several villages . For AP, this dummy variable is taken for the FA instead. We also take a dummy for households residing in the village where the FA and sarpanch both live. The objective is to test if households residing in the village of the sarpanch and FAs get more benefits from the programme. Our qualitative evidence indicates that this is where the core of the support base of the sarpanches lives. Moreover, proximity to the voters makes it easier for the sarpanches to acquire information on who voted for whom. It is also plausible to assume that, for politically ambitious FAs, their village would be the main source of political support. 23 We include a number of explanatory variables to capture the economic, demographic and village level factors. Caste dummies are included as it is a crucial indicator of backwardness and social discrimination. Land holdings, household size and occupational structure are included as these variables can serve as the major indicators of the socio-economic condition of a household. Furthermore, the distance of the household from the GP headquarters is taken as indicator of isolation and seclusion from the GP centre. Moreover, this variable can control for the fact that AP GPs are not as scattered as those in Rajasthan. This can make it very difficult for the FAs to discriminate across the households that live in different villages.
To control for other occupational opportunities for the household apart from MGNREGA, we use the logarithmic value of the reported non-MGNREGA wages received when MGNREGA works are implemented. Further, we incorporate political party dummies in the regression to control for the party locally in power. For AP, a dummy of whether any member of the household or immediate relative is an elected representative of the GP is included as one of the explanatory variables in the econometric exercise. For Rajasthan, we were not able to include this variable because of less variability in the data. In Rajasthan, we have also controlled for woman/SC/ST reserved GPs. For AP, however, we have not used this variable as all the reserved GPs are ruled by the YSR Congress or the Telugu Desam Party (TDP) and hence this variable is already captured by the political variable we incorporated. Since caste composition in a village can determine demand for MGNREGA works, we also control for the proportion of SC or ST in the village where the household resides. Finally, we control for meteorological conditions, as a lack of rainfall can determine a surge in demand for MGNREGA work. We thus include district-level rainfall data for 2012 (proportion of monsoon months from June to September, when deficit rainfall was experienced as against the long-term average in that month) as a demand-side control variable.

(b) Regression Strategy
The first objective of the econometric exercise is to explore if households residing in the village where the sarpanches/FAs reside have a higher probability of getting work under MGNREGA. Since getting work is a dichotomous variable, we apply a logit regression to estimate the determinants of a household working under MGNREGA in 2013-2014. The second and third objectives are to find if these households work for a higher number of days under MGNREGA and earn more in comparison to others. The number of workdays or earnings is censored from below since a substantial proportion of the households did not work, and hence the variables would show a value of zero for these households. Since simple Ordinary Least Squares (OLS) estimates might yield biased estimates, we apply a tobit regression (Long, 1997).
(c) Descriptive Statistics Table 3 gives an indication of how MGNREGA has been implemented in the survey areas. It is found that 41% of the sample households in Rajasthan worked in MGNREGA in 2013-2014. These households worked for 37.5 days and earned Rs. 4027.5 on average from the programme. About 57% of these households are found to reside in the village where the GP sarpanch resides. For AP, 48% of the sampled households worked under MGNREGA in 2013-2014 for 61.5 days on average and earned Rs. 8268.5. Furthermore, 60% of the sampled households are in villages where the FA resides, and 85% of these villages are also where the sarpanch lives. Table 4 shows the regression results for Rajasthan, which contain the estimates for the probability of getting work under MGNREGA in 2013-2014, as well as that for the number of days of work and earnings. The first column gives the odds-ratio. An odds-ratio greater than 1 indicates a positive relationship, which implies a greater chance of achieving the outcome for a particular group in comparison to the reference group. An odds-ratio lower than 1 indicates a negative relationship. We find that households residing in the same village of the sarpanch have significantly higher chances of working under the programme and also of earning more. This indicates that MGNREGA can be seen as an instrument for the sarpanch to get political leverage by allocating the benefits of the programme proportionately more to the households that live in close proximity to him/her. Mukhopadhyay et al. (2015) show similar findings in Rajasthan.  Table 5 shows the estimates from the regression for AP. Analogous to the results from the estimates for Rajasthan, we find households that are located in the same villages as the FAs work significantly more in terms of the number of days and earnings from the programme, though no significant relationship is found in terms of probability of getting work. Interestingly, the findings are true for households residing in villages where both FA and sarpanch live. This probably hints at the fact that in AP, the discretion of choosing households and selective rationing comes through the FAs and partially through sarpanches as well. Of note is the fact that these findings are independent of the socio-economic characteristics of the households, which have been controlled for in the regression.

(d) Results
The empirical results show similar findings as observed from the qualitative narratives discussed in the paper. As indicated, the implementation of MGNREGA in Rajasthan comes through the sarpanches, whose incentives to maximise MGNREGA employment are rather weak. They will therefore be inclined to distribute work in a clientelist fashion. We find that they choose to give priority to their own supporters. In fact, we find households residing in the same village as the sarpanch receive greater benefits from MGNREGA. In AP, where the implementation responsibilities are in the hands of the FAs, we find that, unlike in Rajasthan, households who live in the village where the FA resides do not have higher probability of getting work compared to others. However, the FAs tend to prioritise their supporters in terms of allocating a higher number of workdays to these households compared to others who received work, but unlike in Rajasthan, these individuals do not have a higher probability of getting at least some work.

Discussion and conclusion
This paper analysed the set of incentives that shapes the implementation of the MGNREGA as a result of two different implementation models: in Rajasthan, the implementation is in the hands of an elected official who is accountable to the programme's beneficiaries, whereas in AP the key field implementer is a bureaucrat accountable to the state administration.
Based on a combination of quantitative and qualitative analyses, we find that in the centralised setting (AP), clientelism plays a comparatively less prominent role: MGNREGA work is distributed to those who demand it, although the amount of work allocated varies according to political considerations. In the decentralised setting (Rajasthan), both the allocation of work and the amount of work allocated are distributed in a clientelistic fashion.
The key mechanism that explains these findings relates to the incentives that field implementers have to maximize the generation of MGNREGA employment: the more work is available to be distributed, the more local implementers will enlarge the pool of beneficiaries. The political and economic incentives to maximise the availability of MGNREGA work are similar in the two states and are weaker than it might be assumed: what is crucial is that in the centralised setting, field implementers respond to incentives imposed from above to maximise MGNREGA employment, which in turn results in a fairer distribution of work. Our findings have three main implications. First, along with some well-established predictors of clientelism like poverty and party competition, the institutional set up matters greatly. In the two cases analysed in this paper, it is the centralised setting chosen for the implementation of the MGNREGA that allows for specific policy correctives -like the administrative incentives that we discussed -to be introduced and put in operation.
Secondly, clientelism cannot be treated as a dichotomous variable. The ground reality is more complex and the empirical evidence that we present suggests that clientelism should rather be understood as a continuous variable. In our analysis, the distribution of MGNREGA work is clientelistic in both states, but much less so in AP. This is not only an analytical distinction: less clientelism in AP means that households are not excluded from the programme for political reasons, which makes a big difference on the ground.
Furthermore, we show that clientelism is seldom, if ever, an exclusive relationship between a client and a patron that occurs in a vacuum. Instead, both actors are embedded into a net of social and political relations that inevitably affect and shape their own (clientelistic) relationship. The fact that MGNREGA implementers take into consideration the interests of the big farmers when distributing jobs is a clear example of how a clientelistic relationship is affected by the broader political and social context.
Finally, our analysis shows that democratic decentralisation facilitates clientelism, but also that bureaucracy-led implementation does not completely eliminate it. Moreover, top-down implementation is inherently fragile: the existence of administrative incentives that limit clientelism in AP is crucially dependent on state officials at higher levels who are committed to enforce targets. A few administrative transfers could change the situation radically, altering local implementers' incentives and their effects on clientelistic practice.
As a final normative remark, the findings presented here support the argument that policymakers should pay much more attention to the political context in which policies are implemented. In order to ensure a fair distribution of MGNREGA work and ensure its demand-driven nature, local implementers should be incentivised to generate as much employment as possible. While a key policy correction in this regard is to guarantee enough financial resources, the political context still matters greatly. For instance, establishing a dialogue with farmers' associations in order to ensure continuity in the availability of MGNREGA jobs throughout the year could make a big difference in terms of ensuring that the gap between demand and supply of work is minimised, a key indicator of the success of the policy (Gaiha, Kulkarni, Pandey, & Imai, 2010). Similarly, offering rewards to highly performing GPs offers sarpanches incentives to guarantee the right to work, ultimately translating the rights-based nature of the MGNREGA into reality. Notes 1. In June 2014, the state of AP was bifurcated. The ten northern districts now form the state of Telangana. In this paper, we refer to the unified state. 2. See for example Mansuri & Rao, 2013;Stokes, Dunning, Nazareno, & Brusco, 2013;WDR, 2017, p. 11;Weitz-Shapiro, 2014 3. Elite capture refers to the appropriation of benefits meant for the poor or the public at large by locally powerful elites. 4. Examples of such studies are Weitz-Shapiro (2014) and Heath and Tillin (2018). 5. At least as far as the implementation of the MGNREGA is concerned. In line with other scholars, we take the average number of persondays per household as a key indicator of success (Chopra, 2015). 6. The MGNREGA Act mandates that at least 50% of the funds are spent through the GPs. AP has blatantly violated this provision of the national Act. This was confirmed to us in interviews with Member of Legislative Assembly Mahatma Gandhi National Rural Employment Guarantee Act (MGNREGA), education, and empirical political economy. He primarily works on Development Economics, applied Econometrics and Impact Evaluation.

Silvia Masiero is a lecturer in International Development at the School of Business and Economics, Loughborough
University. Her research interests include the politics of anti-poverty programmes, and the role of Information and Communication Technologies (ICTs) in redesigning social safety nets.